Worked Example

In order to clarify the issu es discussed above, set them in the A bN LP environm ent, and demonstrate the A S-E G flow o f control, a small exam ple is exam ined and the planning process engaged in its production is analysed. The example is taken from (Reed and Long, 1997b), w hich, in (Reed and L ong, 1997d) is discussed in relation to its original textual form (the original argument was taken from the corpus). O nly the propositional structure is exam ined here.

T h e initial situation is one in w hich the hearer is assum ed to have no relevant b eliefs, and the speaker h o ld s a to be true, supported by b, c, and d', c and d are further supported by e and / , respectively. T his structure is show n in Figure 3.7 below .

e f

m . FRAMING A SOLUTION 71

The process is initiated with two goals, BEL ( a g o , a ) , expressing the aim that the hearer, agent ago, should believe the proposition a, and i s _ s a l i e n t ( a g o , a , _ ), expressing the aim that a

should be salient to the hearer. The third parameter on the second goal represents the context in which a proposition is to be made salient; it’s role is explored in more detail in chapter four. The underscore represents the initial situation (corresponding closely to the mutual initial discourse context). It is goals of belief which give rise to structural planning; those of saliency introduce flexibility into the way information is expressed (as discussed in §3.1.6, above).

The communicative goal e e l represents a problem for the classical planning framework, since it is inappropriate to consider it simply as an achievement goal (one which can be satisfied by a single operator). For a BEL is often best satisfied by several operators, i.e. by multiple subarguments (which have been shown to occur with great fi-equency in natural language, (Freeman, 1991), and play a key role in argumentation (Reed and Long, 1997a). However, it is also inappropriate to consider BEL a

maintenance goal, such as the stylistic goals in Hovy’s (1990) system, PAULINE. For it is not the case that BEL goals remain unsatisfied; rather, they are satisfied a number of times and are then considered fulfilled in the classical sense. This iteration problem in discourse planning has generally been approached through the use of some Tor-all’ operation. Maybury (1993), for example, makes explicit use of V in his operator descriptions, and Moore and Paris (1994) introduce a f o r a l l clause. In both papers, however, it is noted that the approach requires explicit, unprincipled modification of the plan language. In the absence of a principled solution, the current work employs a similarly pragmatic notion of universal quantification, to produce a maximal set of support by which to fulfil a b e l goal. (Although UCPOP (Penberthy and Weld, 1992) offers a principled solution to the iteration problem, other features - principally its nonhierarchical nature - make it unsuitable for argument planning).

The first BEL goal, then, is fulfilled by three Modus Ponens operators, viz.MP ( a g o , a , b ) , M P (a g o , a , c ) , and M P(ago, a , d ) , and the IS_SA LIENT goal is fulfilled by a corresponding

MAKE_SALIENT ( a g o , a , _ ). The AS level employs a general heuristic to introduce a soft constraint that the expression of the conclusion - i.e. the m a k e _ s a l i e n t goal - should precede its supports (this heuristic is based upon Blair’s (1838, p.429) distinction between analytic and synthetic argument; the former, where the conclusion is clearly stated at the outset, is to be preferred, ceteris paribus). The result of this first round of planning is the abstract plan in Figure 3.8.

t o : MP ( a g o , a , b ) [ h a r d ] t l : MP( a g o , a , c ) [ s o f t ] t 2 : M P (ago, a , d) [ s o f t ] t 3 : MAKE_SALIENT(ago, a , _ ) [ h a r d ] ( t 3 < tO ) [ s o f t ] ( t 3 < t l ) [ s o f t ] ( t 3 < t 2 ) [ s o f t ]

Figure 3.8 First AS plan with partial order

On the generation of this plan, control passes to the EG level, where a heuristic employing knowledge of the persuasive strength of the various propositions fixes the position of the conclusion. Because the conclusion a is noted in the knowledge base to be rather weak, it is inappropriate to have it-

positioned at the head of the argument: better to have it succeed a strong subargument and then be further supported. In this case, the heuristic fixes its position between to and t l . The first soft ordering suggestion from the AS is thus revoked.

W ith no further structural manipulation (and, in this example, no further stylistic issues) to be considered, the EG returns control. The abstract plan is now complete and undergoes refinement, opening up each of the bodies of the MPs (the m a k e _ s a l i e n t operator, as a primitive, remains in the plan unchanged with respect to content or position). This results in the list of goals in Figure 3.9(a) (in which the partial order and hard/soft indication have been omitted for clarity). Control reverts to the AS level, where further heuristic manipulation of the structure is effected to improve coherency. In particular, a potential problem is detected in the final MP argument (the argument from d). With the position o f the conclusion fixed between the first two subarguments, the coherency of the plan in Figure 3.9(a) is compromised, due to the (relatively) large subargument which intervenes between the expression o f the conclusion, a, and the support leant to it by the third subargument, d. One means of repairing the coherency is to reverse the order of the components in the final subargument such that (d

—> a) is expressed before d itself, thus indicating to the hearer the relevance of d to the conclusion. This reordering is effected in Figure 3.9(b).

(a) (b)

PUSH_TOPIC ( a r g ( b , a ) ) PUSH_TOPIC ( a r g ( b , a ) ) BEL ( h , b ) BEL ( h , b)

IS_SALIENT ( h , b , a r g ( b , a ) ) IS_S ALIENT ( h , b , a r g ( b , a ) ) BEL ( h , b —>a) BEL ( h , b —>a)

IS_SALIENT ( h , b —>a, a r g ( b , a ) ) IS_SA L IEN T ( h , b —»a , a r g ( b , a ) ) POP_TOPIC ( a r g ( b , a ) ) POP_TOPIC ( a r g ( b , a ) )

MAKE_SALIENT ( h , a , _ ) MAKE_SALIENT ( h , a , _ ) PUSH_TOPIC ( a r g ( c , a ) ) PUSH_TOPIC ( a r g ( c , a ) ) BEL ( h , c ) BEL ( h , c )

IS_SALIENT ( h , c , a r g ( c , a ) ) IS_SA LIE N T ( h , c , a r g ( c , a ) ) BEL ( h , c - » a ) BEL ( h , c —>a)

IS_SALIENT ( h , c - » a , a r g ( c , a ) ) IS_SA L IE N T ( h , c - > a , a r g ( c , a ) ) POP_TOPIC ( a r g ( c , a ) ) POP_TOPIC ( a r g ( c , a ) ) PUSH_TOPIC ( a r g ( d , a ) ) PUSH_TOPIC ( a r g ( d , a ) ) BEL ( h , d) BEL ( h , d -> a )

IS_SALIENT ( h , d , a r g ( d , a ) ) IS_SA LIEN T ( h , d —»a, a r g ( d , a ) ) BEL ( h , d - » a ) BEL ( h , d)

IS_SALIENT ( h , d - » a , a r g ( d , a ) ) IS_SA LIEN T ( h , d , a r g ( d , a ) ) POP_TOPIC ( a r g ( d , a ) ) POP_TOPIC ( a r g ( d , a ) )

Figure 3.9 Result of refinement (a) before, and (b) after ordering

Reordering thus occurs in two stages: once at the goal fulfilment phase, and once at the refinement phase. This is necessitated by the requirement that two forms of reordering should be possible: (i) between the supports for a conclusion, and (ii) between supports and their conclusion. This distinction is explained in (Reed and Long, 1997d) by consideration o f orderings within an operator body, such as that of MP (in Figure 3.6, above). Ordering between the supports for a conclusion

m. FRAMING A SOLUTION 7 3

corresponds to arranging the order of the two pairs t l - t 2 and t 3 - t 4 - i.e. whether X precedes X—>P or vice versa. These goals are posted at refinement, necessitating ordering at that time. In contrast, positioning a conclusion corresponds to arranging the order within each of those pairs - i.e. for t l - t 2

whether making X salient should precede or succeed the argument bringing the hearer to believe X, and similarly for t 3 - t 4 , whether making X— salient should precede its support or not. In order to account for placing the conclusion amongst multiple supports, this latter form of ordering needs to occur not between the BEL and i s _ s a l i e n t goal, but between the supports for the b e l and the

IS _SA L IEN T goal. As these supports are generated at goal fulfilment, this second form of ordering has to occur separately.

The small reordering illustrated in Figure 3.9 completes the work of the AS at this level of abstraction: control passes to the EG once more for further reorderings for persuasive effect. In this example, there are none, and the processing for this level of abstraction is complete. The AS then takes as input the set of partially ordered goals in Figure 3.9(b), and performs another round of planning to fulfil those goals. The processing cycle thus comprises five distinct phases, as shown in Figure 3.10, below: first, AS level planning, goal fulfilment, and ordering to maximise coherency; next, EG level processing to effect further reorderings (within the hard constraints laid down at the previous stage); then refinement of the operators selected; next AS reordering with the new body goals; and finally, EG reordering, again within the AS constraints.

AS Planning

EG Planning

Refinement

EG Reordering AS Reordering

Figure 3.10 High level processing cycle.

This double interleaving of AS and EG processing (once with goal fulfilment and once with refinement) ensures that the interplay between coherency-based and persuasion-based considerations are taken into account during both the required ordering phases, endowing the framework with a rich model of the high level generation process.

The broad summary of a round o f planning for a small example given in this section has outlined the functionality of the AS and EG levels, and indicated the means by which the flow of

control passes between them. The next chapter examines the functionality of the AS level in much more detail, exploring the role of the topic manipulators, and enumerating the logical and quasi-logical operators available to the planning process, and the various heuristics for influencing the structure of an argument to improve its coherency.